Qwen3.8-27B-ROCmFPX

ROCmFPX quantized GGUF (7 tiers + vision mmproj) of Qwen/Qwen3.8-27B — official BF16 weights, locally requantized.

⚡ Quick Start

Requires llama.cpp-rocm (this fork supports ROCmFPX; upstream llama.cpp cannot load these files — format note at the end).

# Text only (smallest tier: Q4_FAST ~13.6 GiB)
llama-server -m Qwen3.8-27B_Q4_0_ROCMFP4_FAST.gguf -ngl 99 -fa on --port 8080

# Vision: add --mmproj mmproj-Qwen3.8-27B_BF16.gguf
# All tiers fit on a single 128 GB machine (largest: 26.7 GiB)

🎯 Which tier? (3-second decision)

• Fastest: Q4_0_ROCMFP4_FAST (13.6 GiB)
• Speed + stabler quality: Q4_0_ROCMFP4_STRIX (14.0 GiB — recommended sweet spot)
• Higher-bit reference: Q6_0_ROCMFPX (21.0 GiB)
• Agent / tool-call: Q6_0_ROCMFPX_AGENT (23.6 GiB) / Q8_0_ROCMFPX_AGENT (26.7 GiB)
• Quality first: Q8_0_ROCMFPX (26.3 GiB)

Model Overview

• Base model: Qwen3.8-27B (dense, official weights by Alibaba)
• Quantization: BF16 -> ROCmFP4 / ROCmFPx, 7 variants
• MTP: GGUFs carry an MTP head — enable with llama-server --spec-type draft-mtp (speculative decoding)
• Vision: multimodal — pair with the bundled mmproj (see Usage)

Files

File	Size
Qwen3.8-27B_Q4_0_ROCMFP4_FAST.gguf	13.56 GiB (14,562,235,840 B)
Qwen3.8-27B_Q4_0_ROCMFP4_FAST_COHERENT.gguf	13.90 GiB (14,929,749,440 B)
Qwen3.8-27B_Q4_0_ROCMFP4_STRIX.gguf	13.98 GiB (15,013,963,200 B)
Qwen3.8-27B_Q6_0_ROCMFPX.gguf	20.98 GiB (22,528,382,400 B)
Qwen3.8-27B_Q6_0_ROCMFPX_AGENT.gguf	23.55 GiB (25,283,188,160 B)
Qwen3.8-27B_Q8_0_ROCMFPX.gguf	26.26 GiB (28,193,396,160 B)
Qwen3.8-27B_Q8_0_ROCMFPX_AGENT.gguf	26.69 GiB (28,664,108,480 B)
mmproj-Qwen3.8-27B_BF16.gguf	0.87 GiB (931,145,952 B)

SHA-256 checksums for all files: see sha256.txt.

Quantization Tiers — what's here & how to choose

All files are builds of the same official weights through llama-quantize (llama.cpp-rocm fork); upstream llama.cpp cannot load them (see Format Note). Quick guide:

• Q4_0_ROCMFP4_FAST — smallest & fastest baseline (4.25 bpw). Bulk weights in single-scale FP4.
• Q4_0_ROCMFP4_FAST_COHERENT — FAST + Q6_K token embeddings (better coherence, +0.2 bpw).
• Q4_0_ROCMFP4_STRIX — FAST + Q6_K embeddings + dual-scale attention K/V (quality-biased; same speed class as FAST — the recommended quality/speed sweet spot).
• Q6_0_ROCMFPX / Q8_0_ROCMFPX — higher-bit reference layouts (6.50 / 8.25 bpw). Decode is slower than the Q4 family on this dense model (Q6's unpack overhead is the heaviest).
• Q6_0_ROCMFPX_AGENT / Q8_0_ROCMFPX_AGENT — agent / tool-call oriented routing variants (attention & FFN boosts on sensitive layers).

The full type ladder lives in the llama.cpp-rocm repo (llama-quantize --help).

Benchmarks (llama-bench)

Measured on AMD Strix Halo (gfx1151), llama-bench -ngl 99 -p 256,2048 -n 256 -r 5, llama.cpp-rocm build 11107. (Dense 27B — all parameters active, so tg is inherently lower than MoE models. Each figure is the mean of the last 4 of 5 reps — the first rep carries one-time warm-up cost and is excluded.)

Variant	tg256 (t/s)	pp256 (t/s)	pp2048 (t/s)	Size	bpw
Q4_0_ROCMFP4_FAST	14.05	418.0	398.9	13.6 GiB	4.25
Q4_0_ROCMFP4_FAST_COHERENT	13.99	404.0	389.8	13.9 GiB	~4.45
Q4_0_ROCMFP4_STRIX	14.02	400.2	385.4	14.0 GiB	~4.49
Q6_0_ROCMFPX	9.45	312.1	309.3	21.0 GiB	6.50
Q6_0_ROCMFPX_AGENT	8.70	335.6	330.4	23.6 GiB	6.50
Q8_0_ROCMFPX	7.93	371.8	365.6	26.3 GiB	8.25
Q8_0_ROCMFPX_AGENT	7.83	368.9	363.0	26.7 GiB	8.25

Usage

# Text only:
llama-server -m Qwen3.8-27B_Q4_0_ROCMFP4_FAST.gguf -ngl 99 -fa on --port 8080

# Vision (multimodal; --mmproj enables image input):
llama-server -m Qwen3.8-27B_Q4_0_ROCMFP4_FAST.gguf --mmproj mmproj-Qwen3.8-27B_BF16.gguf -ngl 99 -fa on --port 8080

• Swap -m to any other variant file to run that variant (see Files).
• OpenAI-compatible API served at http://127.0.0.1:8080/v1.
• MTP head: load with llama-server --spec-type draft-mtp for speculative MTP acceleration.

Format Note (ROCmFPX)

ROCmFPX is a quantization format family (llama.cpp-rocm extension: GGML types 100–107 / FTYPE 100–119) — upstream llama.cpp cannot load these files. Load them with llama.cpp-rocm, a llama.cpp fork that supports all upstream GGUF formats plus ROCmFP4/ROCmFPx natively.

License & Disclaimer

• Base model: Qwen3.8-27B (Alibaba) — Apache-2.0
• This quantized version: Apache-2.0 (ArtomYuan)

This is a third-party community quantization (not an official Qwen release), not affiliated with Alibaba / Qwen. Provided for research and learning; use is subject to the base model license (Apache-2.0). No warranty is provided.